Agentic AI Security | AI Field Notes #98

AnalysisThe money is chasing the thing that guards the agents, not just the agents. HiddenLayer, an AI-security firm, raised a $100 million Series B on September 2 led by Delta-v Capital, with Microsoft's M12 fund and Booz Allen Ventures joining, pushing its total to roughly $156 million. The headline product is Agent Harness Security, which watches an AI coding agent at runtime and blocks the prompt injections and tool-abuse tricks that turn a helpful agent into an attacker inside a company's systems. The raise landed the same day two other agent-security deals closed, a sign the defensive layer around autonomous agents is becoming its own market.

AnalysisCoding agents burn money re-reading files. Every time an agent greps a codebase and pulls whole files into context to answer a question, the tokens (the billable units a model reads and writes) pile up. SonarSource, the company behind the SonarQube code-quality tools, shipped Sonar Vortex on September 2 to cut that waste. It keeps a live semantic map of a repository, called SemSitter, that answers navigation queries such as where a function is called or which class owns a method, so the agent carries far less context per turn. SonarSource measured up to 36% lower token cost across Java, Python, JavaScript, C#, and Rust, and ships a five-line hook to wire it into Claude Code.

AnalysisA frontier model got sharply better at writing real software overnight, and the price did not move. Alibaba's Qwen team re-trained its flagship Qwen3.8-Max on September 1 and tagged the build 0902, leaving the 2.4-trillion-parameter model, its 1-million-token context window, and its $2-in, $6-out per-million-token pricing untouched. On TerminalBench 3.0, a test of whether a model can drive a coding agent through a real terminal, the score climbed from 11.3 to 29.0. DeepSWE, a benchmark for fixing real software bugs, went from 56.6 to 69.3. The upgrade arrived as a free in-place swap. Same model page, same bill, better code.

AnalysisAn AI agent that pays your bills without checking in each time is close to real for a billion people. India is preparing a framework, likely to be unveiled next week at the Global Fintech Fest in Mumbai, that would let software agents make small, repeated payments over UPI, the state-backed rails that moved 24.51 billion transactions worth about $314 billion in August, Reuters reported on September 1. It leans on existing UPI tools that delegate payment authority and block funds in advance, capped today near 10,000 rupees, about $105. A person sets caps, rules, and identity checks up front, and the agent spends inside them. The question moves from whether agents can transact to how tightly a person can leash one.

AnalysisAutonomous agents are showing up on both sides of the firewall, as legitimate users and as attackers probing for holes at machine speed. Huskeys, a Tel Aviv startup, raised a $27 million Series A led by Blackstone's venture arm on September 2, at a valuation above $100 million and $35 million raised in total. It sells what it calls Network Edge Security Management, tools that watch the traffic hitting a company's apps and decide which automated requests to allow and which to drop. Huskeys names TikTok, Hugging Face, and the telehealth firm Ro among its customers. The bet is that bot traffic stops being an afterthought and becomes a budget line.

AnalysisFive hundred million dollars for a two-year-old company that automates IT help-desk tickets tells you where security spending is going. Palo Alto Networks agreed to buy Console, sources told TechCrunch, in a cash-and-stock deal announced September 2. Console, founded in 2024 and last valued at $157 million, uses AI agents to investigate and resolve routine IT and security alerts in plain language. Palo Alto will wire it into Cortex, its automated threat-detection platform. The price, more than triple Console's last valuation, shows how hard the incumbents will pay to put an agent between an analyst and the alert queue.

AnalysisThe problem stopped being demand and became parts. Hewlett Packard Enterprise reported $12.2 billion in revenue for its third quarter on September 2, up 34% from a year earlier, with AI system orders up 42% and a record backlog. The awkward number underneath: orders grew three and a half times faster than revenue, because HPE cannot get enough memory and NAND flash, the chips that store data in servers, to build what customers have already bought. It expects the shortage to run into fiscal 2027. Shares fell after hours. A record order book means little when the loading dock is waiting on chips.

AnalysisRobots need to see and feel before they can work, and investors just put $165 million behind one team's version of that. Lyte, founded in 2021 by former Apple sensing engineers, closed a Series C on September 2 at a $1.6 billion valuation led by Maverick Silicon, tripling its worth and bringing total funding to $272 million. It builds the whole perception stack for physical AI: custom silicon, multimodal sensors, and the spatial software a machine uses to know where it is and what moves around it. The valuation is a bet that the hard part of a useful robot is now the senses that link its motors to its models.

AnalysisReal-time graphics took a step toward film-grade lighting, and it only runs on the newest hardware. Nvidia turned on DLSS 5 on September 3, debuting its 3D-Guided Neural Rendering in the basketball game NBA 2K27. Earlier DLSS versions upscaled and filled in frames; this release feeds the game's own geometry and lighting into a neural network that adds effects such as subsurface scattering, the way light passes through skin and ears, and contact shadows that older real-time engines fake. One constraint sits underneath: DLSS 5 is locked to the RTX 50-series (Blackwell) cards, and older GPUs are shut out for now. Nvidia says it will explore bringing it to earlier chips later.

AnalysisThe largest school district in the country just told 600,000 children they cannot use generative AI at school, roughly two-thirds of its enrollment. New York City's education department set a one-year moratorium, announced September 2, that blocks student-facing generative AI, ChatGPT and Claude included, along with companion chatbots, for grades 2K through eight. It pairs the ban with screen-time caps, 30 minutes a day for third through fifth grade and 45 for sixth through eighth, and limits high schoolers to five vetted programs at 45 minutes a week under a teacher's eye. Teachers keep AI for lesson planning. The district wants kids reading before they prompt.

AnalysisThe interesting agent workflows are the boring ones that just run. GitHub published a look on September 2 at a production Copilot setup it calls PR Sous Chef, which scans a repository's open pull requests every 15 minutes, decides which ones actually need a human, and fires targeted Copilot actions only on those. Instead of an agent that reacts to every event or waits for a person to ask, it runs on a clock and stays quiet until a pull request is stuck or ready. It is a small pattern, but it points at where coding agents are heading: scheduled, selective, and mostly invisible.

AnalysisEurope is spending big to train AI on its own soil, and it picked AMD to do it. The EuroHPC Joint Undertaking, the EU's public supercomputing body, signed a 387.8 million euro contract on August 31 for LUMI-AI, a machine to be built in Kajaani, Finland, by France's Bull. It runs on AMD Instinct MI430X accelerators and 256-core EPYC processors, skipping the Nvidia GPUs that dominate almost every other large AI build. EuroHPC expects roughly ten times the AI capacity of the current LUMI system when it comes online in the second half of 2027. The choice is as much about sovereignty as raw speed.